NanoDet

Super fast and lightweight anchor-free object detection model. Real-time on mobile devices.

• ⚡Super lightweight: Model file is only 980KB(INT8) or 1.8MB(FP16).
• ⚡Super fast: 97fps(10.23ms) on mobile ARM CPU.
• 😎Training friendly: Much lower GPU memory cost than other models. Batch-size=80 is available on GTX1060 6G.
• 😎Easy to deploy: Provide C++ implementation with various backends and Android demo based on ncnn inference framework.

---

NEWS!!!

• [2021.08.28] Refactor data processing pipeline and support multi-scale training (#311).

• [2021.05.30] Release ncnn int8 models, and new pre-trained models with ShuffleNetV2-1.5x backbone. Much higher mAP but still realtime(26.8mAP 21.53ms).

• [2021.03.12] Apply the Transformer encoder to NanoDet! Introducing NanoDet-t, which replaces the PAN in NanoDet-m with a TAN(Transformer Attention Net), gets 21.7 mAP(+1.1) on COCO val 2017. Check nanodet-t.yml for more details.

Find more update notes in Update notes.

---

Benchmarks

Model	Resolution	COCO mAP	Latency(ARM 4 Threads)	FLOPS	Params	Model Size
NanoDet-m	320*320	20.6	10.23ms	0.72G	0.95M	1.8MB(FP16) | 980KB(INT8)
NanoDet-m	416*416	23.5	16.44ms	1.2G	0.95M	1.8MB(FP16) | 980KB(INT8)
NanoDet-m-1.5x	320*320	23.5	13.53ms	1.44G	2.08M	3.9MB(FP16) | 2MB(INT8)
NanoDet-m-1.5x	416*416	26.8	21.53ms	2.42G	2.08M	3.9MB(FP16) | 2MB(INT8)
NanoDet-g	416*416	22.9	Not Designed For ARM	4.2G	3.81M	7.7MB(FP16) | 3.6MB(INT8)
YoloV3-Tiny	416*416	16.6	37.6ms	5.62G	8.86M	33.7MB
YoloV4-Tiny	416*416	21.7	32.81ms	6.96G	6.06M	23.0MB

Find more models in Model Zoo

Note:

• Performance is measured on Kirin 980(4xA76+4xA55) ARM CPU based on ncnn. You can test latency on your phone with ncnn_android_benchmark.

• NanoDet mAP(0.5:0.95) is validated on COCO val2017 dataset with no testing time augmentation.

• YOLO mAP refers from Scaled-YOLOv4: Scaling Cross Stage Partial Network.

• NanoDet-g is designed for edge NPU, GPU or TPU with high parallel computing power but low memory bandwidth. It has much lower memory access cost than NanoDet-m.

---

NanoDet is a FCOS-style one-stage anchor-free object detection model which using ATSS for target sampling and using Generalized Focal Loss for classification and box regression. Please refer to these papers for more details.

Fcos: Fully convolutional one-stage object detection

ATSS:Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection

Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

知乎中文介绍 | QQ交流群：908606542 (答案：炼丹)

---

Demo

Android demo

Android demo project is in demo_android_ncnn folder. Please refer to Android demo guide.

Here is a better implementation 👉 ncnn-android-nanodet

NCNN C++ demo

C++ demo based on ncnn is in demo_ncnn folder. Please refer to Cpp demo guide.

MNN demo

Inference using Alibaba's MNN framework is in demo_mnn folder. Including python and cpp inference code. Please refer to MNN demo guide.

OpenVINO demo

Inference using OpenVINO is in demo_openvino folder. Please refer to OpenVINO demo guide.

Web browser demo

https://nihui.github.io/ncnn-webassembly-nanodet/

Pytorch demo

First, install requirements and setup NanoDet following installation guide. Then download COCO pretrain weight from here

👉COCO pretrain weight (Google Drive)

• Inference images

python demo/demo.py image --config CONFIG_PATH --model MODEL_PATH --path IMAGE_PATH

• Inference video

python demo/demo.py video --config CONFIG_PATH --model MODEL_PATH --path VIDEO_PATH

• Inference webcam

python demo/demo.py webcam --config CONFIG_PATH --model MODEL_PATH --camid YOUR_CAMERA_ID

Besides, We provide a notebook here to demonstrate how to make it work with PyTorch.

---

Install

Requirements

• Linux or MacOS
• CUDA >= 10.0
• Python >= 3.6
• Pytorch >= 1.6
• experimental support Windows (Notice: Windows not support distributed training before pytorch1.7)

Step

1. Create a conda virtual environment and then activate it.

 conda create -n nanodet python=3.8 -y
 conda activate nanodet

2. Install pytorch

conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c conda-forge

3. Install requirements

pip install Cython termcolor numpy tensorboard pycocotools matplotlib pyaml opencv-python tqdm pytorch-lightning torchmetrics

4. Setup NanoDet

git clone https://github.com/RangiLyu/nanodet.git
cd nanodet
python setup.py develop

---

Model Zoo

NanoDet supports variety of backbones. Go to the config folder to see the sample training config files.

Model	Backbone	Resolution	COCO mAP	FLOPS	Params	Pre-train weight
NanoDet-m	ShuffleNetV2 1.0x	320*320	20.6	0.72G	0.95M	Download
NanoDet-m-416	ShuffleNetV2 1.0x	416*416	23.5	1.2G	0.95M	Download
NanoDet-m-1.5x	ShuffleNetV2 1.5x	320*320	23.5	1.44G	2.08M	Download
NanoDet-m-1.5x-416	ShuffleNetV2 1.5x	416*416	26.8	2.42G	2.08M	Download
NanoDet-m-0.5x	ShuffleNetV2 0.5x	320*320	13.5	0.3G	0.28M	Download
NanoDet-t (NEW)	ShuffleNetV2 1.0x	320*320	21.7	0.96G	1.36M	Download
NanoDet-g	Custom CSP Net	416*416	22.9	4.2G	3.81M	Download
NanoDet-EfficientLite	EfficientNet-Lite0	320*320	24.7	1.72G	3.11M	Download
NanoDet-EfficientLite	EfficientNet-Lite1	416*416	30.3	4.06G	4.01M	Download
NanoDet-EfficientLite	EfficientNet-Lite2	512*512	32.6	7.12G	4.71M	Download
NanoDet-RepVGG	RepVGG-A0	416*416	27.8	11.3G	6.75M	Download

---

How to Train

1. Prepare dataset

   If your dataset annotations are pascal voc xml format, refer to config/nanodet_custom_xml_dataset.yml

   Or convert your dataset annotations to MS COCO format(COCO annotation format details).

2. Prepare config file

   Copy and modify an example yml config file in config/ folder.

   Change save_path to where you want to save model.

   Change num_classes in model->arch->head.

   Change image path and annotation path in both data->train and data->val.

   Set gpu ids, num workers and batch size in device to fit your device.

   Set total_epochs, lr and lr_schedule according to your dataset and batchsize.

   If you want to modify network, data augmentation or other things, please refer to Config File Detail

3. Start training

   NanoDet is now using pytorch lightning for training.

   For both single-GPU or multiple-GPUs, run:

   python tools/train.py CONFIG_FILE_PATH

   For Windows users, if you have problems with the new lightning trainer, try to use tools/deprecated/train.py

   follow this...

   For single GPU, run

   python tools/deprecated/train.py CONFIG_FILE_PATH

   For multi-GPU, NanoDet using distributed training. (Notice: Windows not support distributed training before pytorch1.7) Please run

   python -m torch.distributed.launch --nproc_per_node=GPU_NUM --master_port 29501 tools/deprecated/train.py CONFIG_FILE_PATH

4. Visualize Logs

   TensorBoard logs are saved in save_dir which you set in config file.

   To visualize tensorboard logs, run:

   cd <YOUR_SAVE_DIR>
   tensorboard --logdir ./

---

How to Deploy

NanoDet provide C++ and Android demo based on ncnn library.

1. Convert model

   To convert NanoDet pytorch model to ncnn, you can choose this way: pytorch->onnx->ncnn

   To export onnx model, run tools/export_onnx.py.

   python tools/export_onnx.py --cfg_path ${CONFIG_PATH} --model_path ${PYTORCH_MODEL_PATH}

   Then using onnx-simplifier to simplify onnx structure.

   python -m onnxsim ${INPUT_ONNX_MODEL} ${OUTPUT_ONNX_MODEL}

   Run onnx2ncnn in ncnn tools to generate ncnn .param and .bin file.

   After that, using ncnnoptimize to optimize ncnn model.

   If you have quentions about converting ncnn model, refer to ncnn wiki. https://github.com/Tencent/ncnn/wiki

2. Run NanoDet model with C++

   Please refer to demo_ncnn.

3. Run NanoDet on Android

   Please refer to android_demo.

---

Thanks

https://github.com/Tencent/ncnn

https://github.com/open-mmlab/mmdetection

https://github.com/implus/GFocal

https://github.com/cmdbug/YOLOv5_NCNN

https://github.com/rbgirshick/yacs